1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * VMware VMCI Driver
4  *
5  * Copyright (C) 2012 VMware, Inc. All rights reserved.
6  */
7 
8 #include <linux/vmw_vmci_defs.h>
9 #include <linux/vmw_vmci_api.h>
10 #include <linux/moduleparam.h>
11 #include <linux/interrupt.h>
12 #include <linux/highmem.h>
13 #include <linux/kernel.h>
14 #include <linux/mm.h>
15 #include <linux/module.h>
16 #include <linux/processor.h>
17 #include <linux/sched.h>
18 #include <linux/slab.h>
19 #include <linux/init.h>
20 #include <linux/pci.h>
21 #include <linux/smp.h>
22 #include <linux/io.h>
23 #include <linux/vmalloc.h>
24 
25 #include "vmci_datagram.h"
26 #include "vmci_doorbell.h"
27 #include "vmci_context.h"
28 #include "vmci_driver.h"
29 #include "vmci_event.h"
30 
31 #define PCI_DEVICE_ID_VMWARE_VMCI	0x0740
32 
33 #define VMCI_UTIL_NUM_RESOURCES 1
34 
35 /*
36  * Datagram buffers for DMA send/receive must accommodate at least
37  * a maximum sized datagram and the header.
38  */
39 #define VMCI_DMA_DG_BUFFER_SIZE (VMCI_MAX_DG_SIZE + PAGE_SIZE)
40 
41 static bool vmci_disable_msi;
42 module_param_named(disable_msi, vmci_disable_msi, bool, 0);
43 MODULE_PARM_DESC(disable_msi, "Disable MSI use in driver - (default=0)");
44 
45 static bool vmci_disable_msix;
46 module_param_named(disable_msix, vmci_disable_msix, bool, 0);
47 MODULE_PARM_DESC(disable_msix, "Disable MSI-X use in driver - (default=0)");
48 
49 static u32 ctx_update_sub_id = VMCI_INVALID_ID;
50 static u32 vm_context_id = VMCI_INVALID_ID;
51 
52 struct vmci_guest_device {
53 	struct device *dev;	/* PCI device we are attached to */
54 	void __iomem *iobase;
55 	void __iomem *mmio_base;
56 
57 	bool exclusive_vectors;
58 
59 	struct tasklet_struct datagram_tasklet;
60 	struct tasklet_struct bm_tasklet;
61 	struct wait_queue_head inout_wq;
62 
63 	void *data_buffer;
64 	dma_addr_t data_buffer_base;
65 	void *tx_buffer;
66 	dma_addr_t tx_buffer_base;
67 	void *notification_bitmap;
68 	dma_addr_t notification_base;
69 };
70 
71 static bool use_ppn64;
72 
73 bool vmci_use_ppn64(void)
74 {
75 	return use_ppn64;
76 }
77 
78 /* vmci_dev singleton device and supporting data*/
79 struct pci_dev *vmci_pdev;
80 static struct vmci_guest_device *vmci_dev_g;
81 static DEFINE_SPINLOCK(vmci_dev_spinlock);
82 
83 static atomic_t vmci_num_guest_devices = ATOMIC_INIT(0);
84 
85 bool vmci_guest_code_active(void)
86 {
87 	return atomic_read(&vmci_num_guest_devices) != 0;
88 }
89 
90 u32 vmci_get_vm_context_id(void)
91 {
92 	if (vm_context_id == VMCI_INVALID_ID) {
93 		struct vmci_datagram get_cid_msg;
94 		get_cid_msg.dst =
95 		    vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
96 				     VMCI_GET_CONTEXT_ID);
97 		get_cid_msg.src = VMCI_ANON_SRC_HANDLE;
98 		get_cid_msg.payload_size = 0;
99 		vm_context_id = vmci_send_datagram(&get_cid_msg);
100 	}
101 	return vm_context_id;
102 }
103 
104 static unsigned int vmci_read_reg(struct vmci_guest_device *dev, u32 reg)
105 {
106 	if (dev->mmio_base != NULL)
107 		return readl(dev->mmio_base + reg);
108 	return ioread32(dev->iobase + reg);
109 }
110 
111 static void vmci_write_reg(struct vmci_guest_device *dev, u32 val, u32 reg)
112 {
113 	if (dev->mmio_base != NULL)
114 		writel(val, dev->mmio_base + reg);
115 	else
116 		iowrite32(val, dev->iobase + reg);
117 }
118 
119 static void vmci_read_data(struct vmci_guest_device *vmci_dev,
120 			   void *dest, size_t size)
121 {
122 	if (vmci_dev->mmio_base == NULL)
123 		ioread8_rep(vmci_dev->iobase + VMCI_DATA_IN_ADDR,
124 			    dest, size);
125 	else {
126 		/*
127 		 * For DMA datagrams, the data_buffer will contain the header on the
128 		 * first page, followed by the incoming datagram(s) on the following
129 		 * pages. The header uses an S/G element immediately following the
130 		 * header on the first page to point to the data area.
131 		 */
132 		struct vmci_data_in_out_header *buffer_header = vmci_dev->data_buffer;
133 		struct vmci_sg_elem *sg_array = (struct vmci_sg_elem *)(buffer_header + 1);
134 		size_t buffer_offset = dest - vmci_dev->data_buffer;
135 
136 		buffer_header->opcode = 1;
137 		buffer_header->size = 1;
138 		buffer_header->busy = 0;
139 		sg_array[0].addr = vmci_dev->data_buffer_base + buffer_offset;
140 		sg_array[0].size = size;
141 
142 		vmci_write_reg(vmci_dev, lower_32_bits(vmci_dev->data_buffer_base),
143 			       VMCI_DATA_IN_LOW_ADDR);
144 
145 		wait_event(vmci_dev->inout_wq, buffer_header->busy == 1);
146 	}
147 }
148 
149 static int vmci_write_data(struct vmci_guest_device *dev,
150 			   struct vmci_datagram *dg)
151 {
152 	int result;
153 
154 	if (dev->mmio_base != NULL) {
155 		struct vmci_data_in_out_header *buffer_header = dev->tx_buffer;
156 		u8 *dg_out_buffer = (u8 *)(buffer_header + 1);
157 
158 		if (VMCI_DG_SIZE(dg) > VMCI_MAX_DG_SIZE)
159 			return VMCI_ERROR_INVALID_ARGS;
160 
161 		/*
162 		 * Initialize send buffer with outgoing datagram
163 		 * and set up header for inline data. Device will
164 		 * not access buffer asynchronously - only after
165 		 * the write to VMCI_DATA_OUT_LOW_ADDR.
166 		 */
167 		memcpy(dg_out_buffer, dg, VMCI_DG_SIZE(dg));
168 		buffer_header->opcode = 0;
169 		buffer_header->size = VMCI_DG_SIZE(dg);
170 		buffer_header->busy = 1;
171 
172 		vmci_write_reg(dev, lower_32_bits(dev->tx_buffer_base),
173 			       VMCI_DATA_OUT_LOW_ADDR);
174 
175 		/* Caller holds a spinlock, so cannot block. */
176 		spin_until_cond(buffer_header->busy == 0);
177 
178 		result = vmci_read_reg(vmci_dev_g, VMCI_RESULT_LOW_ADDR);
179 		if (result == VMCI_SUCCESS)
180 			result = (int)buffer_header->result;
181 	} else {
182 		iowrite8_rep(dev->iobase + VMCI_DATA_OUT_ADDR,
183 			     dg, VMCI_DG_SIZE(dg));
184 		result = vmci_read_reg(vmci_dev_g, VMCI_RESULT_LOW_ADDR);
185 	}
186 
187 	return result;
188 }
189 
190 /*
191  * VM to hypervisor call mechanism. We use the standard VMware naming
192  * convention since shared code is calling this function as well.
193  */
194 int vmci_send_datagram(struct vmci_datagram *dg)
195 {
196 	unsigned long flags;
197 	int result;
198 
199 	/* Check args. */
200 	if (dg == NULL)
201 		return VMCI_ERROR_INVALID_ARGS;
202 
203 	/*
204 	 * Need to acquire spinlock on the device because the datagram
205 	 * data may be spread over multiple pages and the monitor may
206 	 * interleave device user rpc calls from multiple
207 	 * VCPUs. Acquiring the spinlock precludes that
208 	 * possibility. Disabling interrupts to avoid incoming
209 	 * datagrams during a "rep out" and possibly landing up in
210 	 * this function.
211 	 */
212 	spin_lock_irqsave(&vmci_dev_spinlock, flags);
213 
214 	if (vmci_dev_g) {
215 		vmci_write_data(vmci_dev_g, dg);
216 		result = vmci_read_reg(vmci_dev_g, VMCI_RESULT_LOW_ADDR);
217 	} else {
218 		result = VMCI_ERROR_UNAVAILABLE;
219 	}
220 
221 	spin_unlock_irqrestore(&vmci_dev_spinlock, flags);
222 
223 	return result;
224 }
225 EXPORT_SYMBOL_GPL(vmci_send_datagram);
226 
227 /*
228  * Gets called with the new context id if updated or resumed.
229  * Context id.
230  */
231 static void vmci_guest_cid_update(u32 sub_id,
232 				  const struct vmci_event_data *event_data,
233 				  void *client_data)
234 {
235 	const struct vmci_event_payld_ctx *ev_payload =
236 				vmci_event_data_const_payload(event_data);
237 
238 	if (sub_id != ctx_update_sub_id) {
239 		pr_devel("Invalid subscriber (ID=0x%x)\n", sub_id);
240 		return;
241 	}
242 
243 	if (!event_data || ev_payload->context_id == VMCI_INVALID_ID) {
244 		pr_devel("Invalid event data\n");
245 		return;
246 	}
247 
248 	pr_devel("Updating context from (ID=0x%x) to (ID=0x%x) on event (type=%d)\n",
249 		 vm_context_id, ev_payload->context_id, event_data->event);
250 
251 	vm_context_id = ev_payload->context_id;
252 }
253 
254 /*
255  * Verify that the host supports the hypercalls we need. If it does not,
256  * try to find fallback hypercalls and use those instead.  Returns
257  * true if required hypercalls (or fallback hypercalls) are
258  * supported by the host, false otherwise.
259  */
260 static int vmci_check_host_caps(struct pci_dev *pdev)
261 {
262 	bool result;
263 	struct vmci_resource_query_msg *msg;
264 	u32 msg_size = sizeof(struct vmci_resource_query_hdr) +
265 				VMCI_UTIL_NUM_RESOURCES * sizeof(u32);
266 	struct vmci_datagram *check_msg;
267 
268 	check_msg = kzalloc(msg_size, GFP_KERNEL);
269 	if (!check_msg) {
270 		dev_err(&pdev->dev, "%s: Insufficient memory\n", __func__);
271 		return -ENOMEM;
272 	}
273 
274 	check_msg->dst = vmci_make_handle(VMCI_HYPERVISOR_CONTEXT_ID,
275 					  VMCI_RESOURCES_QUERY);
276 	check_msg->src = VMCI_ANON_SRC_HANDLE;
277 	check_msg->payload_size = msg_size - VMCI_DG_HEADERSIZE;
278 	msg = (struct vmci_resource_query_msg *)VMCI_DG_PAYLOAD(check_msg);
279 
280 	msg->num_resources = VMCI_UTIL_NUM_RESOURCES;
281 	msg->resources[0] = VMCI_GET_CONTEXT_ID;
282 
283 	/* Checks that hyper calls are supported */
284 	result = vmci_send_datagram(check_msg) == 0x01;
285 	kfree(check_msg);
286 
287 	dev_dbg(&pdev->dev, "%s: Host capability check: %s\n",
288 		__func__, result ? "PASSED" : "FAILED");
289 
290 	/* We need the vector. There are no fallbacks. */
291 	return result ? 0 : -ENXIO;
292 }
293 
294 /*
295  * Reads datagrams from the device and dispatches them. For IO port
296  * based access to the device, we always start reading datagrams into
297  * only the first page of the datagram buffer. If the datagrams don't
298  * fit into one page, we use the maximum datagram buffer size for the
299  * remainder of the invocation. This is a simple heuristic for not
300  * penalizing small datagrams. For DMA-based datagrams, we always
301  * use the maximum datagram buffer size, since there is no performance
302  * penalty for doing so.
303  *
304  * This function assumes that it has exclusive access to the data
305  * in register(s) for the duration of the call.
306  */
307 static void vmci_dispatch_dgs(unsigned long data)
308 {
309 	struct vmci_guest_device *vmci_dev = (struct vmci_guest_device *)data;
310 	u8 *dg_in_buffer = vmci_dev->data_buffer;
311 	struct vmci_datagram *dg;
312 	size_t dg_in_buffer_size = VMCI_MAX_DG_SIZE;
313 	size_t current_dg_in_buffer_size;
314 	size_t remaining_bytes;
315 	bool is_io_port = vmci_dev->mmio_base == NULL;
316 
317 	BUILD_BUG_ON(VMCI_MAX_DG_SIZE < PAGE_SIZE);
318 
319 	if (!is_io_port) {
320 		/* For mmio, the first page is used for the header. */
321 		dg_in_buffer += PAGE_SIZE;
322 
323 		/*
324 		 * For DMA-based datagram operations, there is no performance
325 		 * penalty for reading the maximum buffer size.
326 		 */
327 		current_dg_in_buffer_size = VMCI_MAX_DG_SIZE;
328 	} else {
329 		current_dg_in_buffer_size = PAGE_SIZE;
330 	}
331 	vmci_read_data(vmci_dev, dg_in_buffer, current_dg_in_buffer_size);
332 	dg = (struct vmci_datagram *)dg_in_buffer;
333 	remaining_bytes = current_dg_in_buffer_size;
334 
335 	/*
336 	 * Read through the buffer until an invalid datagram header is
337 	 * encountered. The exit condition for datagrams read through
338 	 * VMCI_DATA_IN_ADDR is a bit more complicated, since a datagram
339 	 * can start on any page boundary in the buffer.
340 	 */
341 	while (dg->dst.resource != VMCI_INVALID_ID ||
342 	       (is_io_port && remaining_bytes > PAGE_SIZE)) {
343 		unsigned dg_in_size;
344 
345 		/*
346 		 * If using VMCI_DATA_IN_ADDR, skip to the next page
347 		 * as a datagram can start on any page boundary.
348 		 */
349 		if (dg->dst.resource == VMCI_INVALID_ID) {
350 			dg = (struct vmci_datagram *)roundup(
351 				(uintptr_t)dg + 1, PAGE_SIZE);
352 			remaining_bytes =
353 				(size_t)(dg_in_buffer +
354 					 current_dg_in_buffer_size -
355 					 (u8 *)dg);
356 			continue;
357 		}
358 
359 		dg_in_size = VMCI_DG_SIZE_ALIGNED(dg);
360 
361 		if (dg_in_size <= dg_in_buffer_size) {
362 			int result;
363 
364 			/*
365 			 * If the remaining bytes in the datagram
366 			 * buffer doesn't contain the complete
367 			 * datagram, we first make sure we have enough
368 			 * room for it and then we read the reminder
369 			 * of the datagram and possibly any following
370 			 * datagrams.
371 			 */
372 			if (dg_in_size > remaining_bytes) {
373 				if (remaining_bytes !=
374 				    current_dg_in_buffer_size) {
375 
376 					/*
377 					 * We move the partial
378 					 * datagram to the front and
379 					 * read the reminder of the
380 					 * datagram and possibly
381 					 * following calls into the
382 					 * following bytes.
383 					 */
384 					memmove(dg_in_buffer, dg_in_buffer +
385 						current_dg_in_buffer_size -
386 						remaining_bytes,
387 						remaining_bytes);
388 					dg = (struct vmci_datagram *)
389 					    dg_in_buffer;
390 				}
391 
392 				if (current_dg_in_buffer_size !=
393 				    dg_in_buffer_size)
394 					current_dg_in_buffer_size =
395 					    dg_in_buffer_size;
396 
397 				vmci_read_data(vmci_dev,
398 					       dg_in_buffer +
399 						remaining_bytes,
400 					       current_dg_in_buffer_size -
401 						remaining_bytes);
402 			}
403 
404 			/*
405 			 * We special case event datagrams from the
406 			 * hypervisor.
407 			 */
408 			if (dg->src.context == VMCI_HYPERVISOR_CONTEXT_ID &&
409 			    dg->dst.resource == VMCI_EVENT_HANDLER) {
410 				result = vmci_event_dispatch(dg);
411 			} else {
412 				result = vmci_datagram_invoke_guest_handler(dg);
413 			}
414 			if (result < VMCI_SUCCESS)
415 				dev_dbg(vmci_dev->dev,
416 					"Datagram with resource (ID=0x%x) failed (err=%d)\n",
417 					 dg->dst.resource, result);
418 
419 			/* On to the next datagram. */
420 			dg = (struct vmci_datagram *)((u8 *)dg +
421 						      dg_in_size);
422 		} else {
423 			size_t bytes_to_skip;
424 
425 			/*
426 			 * Datagram doesn't fit in datagram buffer of maximal
427 			 * size. We drop it.
428 			 */
429 			dev_dbg(vmci_dev->dev,
430 				"Failed to receive datagram (size=%u bytes)\n",
431 				 dg_in_size);
432 
433 			bytes_to_skip = dg_in_size - remaining_bytes;
434 			if (current_dg_in_buffer_size != dg_in_buffer_size)
435 				current_dg_in_buffer_size = dg_in_buffer_size;
436 
437 			for (;;) {
438 				vmci_read_data(vmci_dev, dg_in_buffer,
439 					       current_dg_in_buffer_size);
440 				if (bytes_to_skip <= current_dg_in_buffer_size)
441 					break;
442 
443 				bytes_to_skip -= current_dg_in_buffer_size;
444 			}
445 			dg = (struct vmci_datagram *)(dg_in_buffer +
446 						      bytes_to_skip);
447 		}
448 
449 		remaining_bytes =
450 		    (size_t) (dg_in_buffer + current_dg_in_buffer_size -
451 			      (u8 *)dg);
452 
453 		if (remaining_bytes < VMCI_DG_HEADERSIZE) {
454 			/* Get the next batch of datagrams. */
455 
456 			vmci_read_data(vmci_dev, dg_in_buffer,
457 				    current_dg_in_buffer_size);
458 			dg = (struct vmci_datagram *)dg_in_buffer;
459 			remaining_bytes = current_dg_in_buffer_size;
460 		}
461 	}
462 }
463 
464 /*
465  * Scans the notification bitmap for raised flags, clears them
466  * and handles the notifications.
467  */
468 static void vmci_process_bitmap(unsigned long data)
469 {
470 	struct vmci_guest_device *dev = (struct vmci_guest_device *)data;
471 
472 	if (!dev->notification_bitmap) {
473 		dev_dbg(dev->dev, "No bitmap present in %s\n", __func__);
474 		return;
475 	}
476 
477 	vmci_dbell_scan_notification_entries(dev->notification_bitmap);
478 }
479 
480 /*
481  * Interrupt handler for legacy or MSI interrupt, or for first MSI-X
482  * interrupt (vector VMCI_INTR_DATAGRAM).
483  */
484 static irqreturn_t vmci_interrupt(int irq, void *_dev)
485 {
486 	struct vmci_guest_device *dev = _dev;
487 
488 	/*
489 	 * If we are using MSI-X with exclusive vectors then we simply schedule
490 	 * the datagram tasklet, since we know the interrupt was meant for us.
491 	 * Otherwise we must read the ICR to determine what to do.
492 	 */
493 
494 	if (dev->exclusive_vectors) {
495 		tasklet_schedule(&dev->datagram_tasklet);
496 	} else {
497 		unsigned int icr;
498 
499 		/* Acknowledge interrupt and determine what needs doing. */
500 		icr = vmci_read_reg(dev, VMCI_ICR_ADDR);
501 		if (icr == 0 || icr == ~0)
502 			return IRQ_NONE;
503 
504 		if (icr & VMCI_ICR_DATAGRAM) {
505 			tasklet_schedule(&dev->datagram_tasklet);
506 			icr &= ~VMCI_ICR_DATAGRAM;
507 		}
508 
509 		if (icr & VMCI_ICR_NOTIFICATION) {
510 			tasklet_schedule(&dev->bm_tasklet);
511 			icr &= ~VMCI_ICR_NOTIFICATION;
512 		}
513 
514 
515 		if (icr & VMCI_ICR_DMA_DATAGRAM) {
516 			wake_up_all(&dev->inout_wq);
517 			icr &= ~VMCI_ICR_DMA_DATAGRAM;
518 		}
519 
520 		if (icr != 0)
521 			dev_warn(dev->dev,
522 				 "Ignoring unknown interrupt cause (%d)\n",
523 				 icr);
524 	}
525 
526 	return IRQ_HANDLED;
527 }
528 
529 /*
530  * Interrupt handler for MSI-X interrupt vector VMCI_INTR_NOTIFICATION,
531  * which is for the notification bitmap.  Will only get called if we are
532  * using MSI-X with exclusive vectors.
533  */
534 static irqreturn_t vmci_interrupt_bm(int irq, void *_dev)
535 {
536 	struct vmci_guest_device *dev = _dev;
537 
538 	/* For MSI-X we can just assume it was meant for us. */
539 	tasklet_schedule(&dev->bm_tasklet);
540 
541 	return IRQ_HANDLED;
542 }
543 
544 /*
545  * Interrupt handler for MSI-X interrupt vector VMCI_INTR_DMA_DATAGRAM,
546  * which is for the completion of a DMA datagram send or receive operation.
547  * Will only get called if we are using MSI-X with exclusive vectors.
548  */
549 static irqreturn_t vmci_interrupt_dma_datagram(int irq, void *_dev)
550 {
551 	struct vmci_guest_device *dev = _dev;
552 
553 	wake_up_all(&dev->inout_wq);
554 
555 	return IRQ_HANDLED;
556 }
557 
558 static void vmci_free_dg_buffers(struct vmci_guest_device *vmci_dev)
559 {
560 	if (vmci_dev->mmio_base != NULL) {
561 		if (vmci_dev->tx_buffer != NULL)
562 			dma_free_coherent(vmci_dev->dev,
563 					  VMCI_DMA_DG_BUFFER_SIZE,
564 					  vmci_dev->tx_buffer,
565 					  vmci_dev->tx_buffer_base);
566 		if (vmci_dev->data_buffer != NULL)
567 			dma_free_coherent(vmci_dev->dev,
568 					  VMCI_DMA_DG_BUFFER_SIZE,
569 					  vmci_dev->data_buffer,
570 					  vmci_dev->data_buffer_base);
571 	} else {
572 		vfree(vmci_dev->data_buffer);
573 	}
574 }
575 
576 /*
577  * Most of the initialization at module load time is done here.
578  */
579 static int vmci_guest_probe_device(struct pci_dev *pdev,
580 				   const struct pci_device_id *id)
581 {
582 	struct vmci_guest_device *vmci_dev;
583 	void __iomem *iobase = NULL;
584 	void __iomem *mmio_base = NULL;
585 	unsigned int num_irq_vectors;
586 	unsigned int capabilities;
587 	unsigned int caps_in_use;
588 	unsigned long cmd;
589 	int vmci_err;
590 	int error;
591 
592 	dev_dbg(&pdev->dev, "Probing for vmci/PCI guest device\n");
593 
594 	error = pcim_enable_device(pdev);
595 	if (error) {
596 		dev_err(&pdev->dev,
597 			"Failed to enable VMCI device: %d\n", error);
598 		return error;
599 	}
600 
601 	/*
602 	 * The VMCI device with mmio access to registers requests 256KB
603 	 * for BAR1. If present, driver will use new VMCI device
604 	 * functionality for register access and datagram send/recv.
605 	 */
606 
607 	if (pci_resource_len(pdev, 1) == VMCI_WITH_MMIO_ACCESS_BAR_SIZE) {
608 		dev_info(&pdev->dev, "MMIO register access is available\n");
609 		mmio_base = pci_iomap_range(pdev, 1, VMCI_MMIO_ACCESS_OFFSET,
610 					    VMCI_MMIO_ACCESS_SIZE);
611 		/* If the map fails, we fall back to IOIO access. */
612 		if (!mmio_base)
613 			dev_warn(&pdev->dev, "Failed to map MMIO register access\n");
614 	}
615 
616 	if (!mmio_base) {
617 		error = pcim_iomap_regions(pdev, BIT(0), KBUILD_MODNAME);
618 		if (error) {
619 			dev_err(&pdev->dev, "Failed to reserve/map IO regions\n");
620 			return error;
621 		}
622 		iobase = pcim_iomap_table(pdev)[0];
623 	}
624 
625 	vmci_dev = devm_kzalloc(&pdev->dev, sizeof(*vmci_dev), GFP_KERNEL);
626 	if (!vmci_dev) {
627 		dev_err(&pdev->dev,
628 			"Can't allocate memory for VMCI device\n");
629 		return -ENOMEM;
630 	}
631 
632 	vmci_dev->dev = &pdev->dev;
633 	vmci_dev->exclusive_vectors = false;
634 	vmci_dev->iobase = iobase;
635 	vmci_dev->mmio_base = mmio_base;
636 
637 	tasklet_init(&vmci_dev->datagram_tasklet,
638 		     vmci_dispatch_dgs, (unsigned long)vmci_dev);
639 	tasklet_init(&vmci_dev->bm_tasklet,
640 		     vmci_process_bitmap, (unsigned long)vmci_dev);
641 	init_waitqueue_head(&vmci_dev->inout_wq);
642 
643 	if (mmio_base != NULL) {
644 		vmci_dev->tx_buffer = dma_alloc_coherent(&pdev->dev, VMCI_DMA_DG_BUFFER_SIZE,
645 							 &vmci_dev->tx_buffer_base,
646 							 GFP_KERNEL);
647 		if (!vmci_dev->tx_buffer) {
648 			dev_err(&pdev->dev,
649 				"Can't allocate memory for datagram tx buffer\n");
650 			return -ENOMEM;
651 		}
652 
653 		vmci_dev->data_buffer = dma_alloc_coherent(&pdev->dev, VMCI_DMA_DG_BUFFER_SIZE,
654 							   &vmci_dev->data_buffer_base,
655 							   GFP_KERNEL);
656 	} else {
657 		vmci_dev->data_buffer = vmalloc(VMCI_MAX_DG_SIZE);
658 	}
659 	if (!vmci_dev->data_buffer) {
660 		dev_err(&pdev->dev,
661 			"Can't allocate memory for datagram buffer\n");
662 		error = -ENOMEM;
663 		goto err_free_data_buffers;
664 	}
665 
666 	pci_set_master(pdev);	/* To enable queue_pair functionality. */
667 
668 	/*
669 	 * Verify that the VMCI Device supports the capabilities that
670 	 * we need. If the device is missing capabilities that we would
671 	 * like to use, check for fallback capabilities and use those
672 	 * instead (so we can run a new VM on old hosts). Fail the load if
673 	 * a required capability is missing and there is no fallback.
674 	 *
675 	 * Right now, we need datagrams. There are no fallbacks.
676 	 */
677 	capabilities = vmci_read_reg(vmci_dev, VMCI_CAPS_ADDR);
678 	if (!(capabilities & VMCI_CAPS_DATAGRAM)) {
679 		dev_err(&pdev->dev, "Device does not support datagrams\n");
680 		error = -ENXIO;
681 		goto err_free_data_buffers;
682 	}
683 	caps_in_use = VMCI_CAPS_DATAGRAM;
684 
685 	/*
686 	 * Use 64-bit PPNs if the device supports.
687 	 *
688 	 * There is no check for the return value of dma_set_mask_and_coherent
689 	 * since this driver can handle the default mask values if
690 	 * dma_set_mask_and_coherent fails.
691 	 */
692 	if (capabilities & VMCI_CAPS_PPN64) {
693 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
694 		use_ppn64 = true;
695 		caps_in_use |= VMCI_CAPS_PPN64;
696 	} else {
697 		dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(44));
698 		use_ppn64 = false;
699 	}
700 
701 	/*
702 	 * If the hardware supports notifications, we will use that as
703 	 * well.
704 	 */
705 	if (capabilities & VMCI_CAPS_NOTIFICATIONS) {
706 		vmci_dev->notification_bitmap = dma_alloc_coherent(
707 			&pdev->dev, PAGE_SIZE, &vmci_dev->notification_base,
708 			GFP_KERNEL);
709 		if (!vmci_dev->notification_bitmap) {
710 			dev_warn(&pdev->dev,
711 				 "Unable to allocate notification bitmap\n");
712 		} else {
713 			memset(vmci_dev->notification_bitmap, 0, PAGE_SIZE);
714 			caps_in_use |= VMCI_CAPS_NOTIFICATIONS;
715 		}
716 	}
717 
718 	if (mmio_base != NULL) {
719 		if (capabilities & VMCI_CAPS_DMA_DATAGRAM) {
720 			caps_in_use |= VMCI_CAPS_DMA_DATAGRAM;
721 		} else {
722 			dev_err(&pdev->dev,
723 				"Missing capability: VMCI_CAPS_DMA_DATAGRAM\n");
724 			error = -ENXIO;
725 			goto err_free_data_buffers;
726 		}
727 	}
728 
729 	dev_info(&pdev->dev, "Using capabilities 0x%x\n", caps_in_use);
730 
731 	/* Let the host know which capabilities we intend to use. */
732 	vmci_write_reg(vmci_dev, caps_in_use, VMCI_CAPS_ADDR);
733 
734 	if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM) {
735 		/* Let the device know the size for pages passed down. */
736 		vmci_write_reg(vmci_dev, PAGE_SHIFT, VMCI_GUEST_PAGE_SHIFT);
737 
738 		/* Configure the high order parts of the data in/out buffers. */
739 		vmci_write_reg(vmci_dev, upper_32_bits(vmci_dev->data_buffer_base),
740 			       VMCI_DATA_IN_HIGH_ADDR);
741 		vmci_write_reg(vmci_dev, upper_32_bits(vmci_dev->tx_buffer_base),
742 			       VMCI_DATA_OUT_HIGH_ADDR);
743 	}
744 
745 	/* Set up global device so that we can start sending datagrams */
746 	spin_lock_irq(&vmci_dev_spinlock);
747 	vmci_dev_g = vmci_dev;
748 	vmci_pdev = pdev;
749 	spin_unlock_irq(&vmci_dev_spinlock);
750 
751 	/*
752 	 * Register notification bitmap with device if that capability is
753 	 * used.
754 	 */
755 	if (caps_in_use & VMCI_CAPS_NOTIFICATIONS) {
756 		unsigned long bitmap_ppn =
757 			vmci_dev->notification_base >> PAGE_SHIFT;
758 		if (!vmci_dbell_register_notification_bitmap(bitmap_ppn)) {
759 			dev_warn(&pdev->dev,
760 				 "VMCI device unable to register notification bitmap with PPN 0x%lx\n",
761 				 bitmap_ppn);
762 			error = -ENXIO;
763 			goto err_remove_vmci_dev_g;
764 		}
765 	}
766 
767 	/* Check host capabilities. */
768 	error = vmci_check_host_caps(pdev);
769 	if (error)
770 		goto err_remove_bitmap;
771 
772 	/* Enable device. */
773 
774 	/*
775 	 * We subscribe to the VMCI_EVENT_CTX_ID_UPDATE here so we can
776 	 * update the internal context id when needed.
777 	 */
778 	vmci_err = vmci_event_subscribe(VMCI_EVENT_CTX_ID_UPDATE,
779 					vmci_guest_cid_update, NULL,
780 					&ctx_update_sub_id);
781 	if (vmci_err < VMCI_SUCCESS)
782 		dev_warn(&pdev->dev,
783 			 "Failed to subscribe to event (type=%d): %d\n",
784 			 VMCI_EVENT_CTX_ID_UPDATE, vmci_err);
785 
786 	/*
787 	 * Enable interrupts.  Try MSI-X first, then MSI, and then fallback on
788 	 * legacy interrupts.
789 	 */
790 	if (vmci_dev->mmio_base != NULL)
791 		num_irq_vectors = VMCI_MAX_INTRS;
792 	else
793 		num_irq_vectors = VMCI_MAX_INTRS_NOTIFICATION;
794 	error = pci_alloc_irq_vectors(pdev, num_irq_vectors, num_irq_vectors,
795 				      PCI_IRQ_MSIX);
796 	if (error < 0) {
797 		error = pci_alloc_irq_vectors(pdev, 1, 1,
798 				PCI_IRQ_MSIX | PCI_IRQ_MSI | PCI_IRQ_LEGACY);
799 		if (error < 0)
800 			goto err_remove_bitmap;
801 	} else {
802 		vmci_dev->exclusive_vectors = true;
803 	}
804 
805 	/*
806 	 * Request IRQ for legacy or MSI interrupts, or for first
807 	 * MSI-X vector.
808 	 */
809 	error = request_irq(pci_irq_vector(pdev, 0), vmci_interrupt,
810 			    IRQF_SHARED, KBUILD_MODNAME, vmci_dev);
811 	if (error) {
812 		dev_err(&pdev->dev, "Irq %u in use: %d\n",
813 			pci_irq_vector(pdev, 0), error);
814 		goto err_disable_msi;
815 	}
816 
817 	/*
818 	 * For MSI-X with exclusive vectors we need to request an
819 	 * interrupt for each vector so that we get a separate
820 	 * interrupt handler routine.  This allows us to distinguish
821 	 * between the vectors.
822 	 */
823 	if (vmci_dev->exclusive_vectors) {
824 		error = request_irq(pci_irq_vector(pdev, 1),
825 				    vmci_interrupt_bm, 0, KBUILD_MODNAME,
826 				    vmci_dev);
827 		if (error) {
828 			dev_err(&pdev->dev,
829 				"Failed to allocate irq %u: %d\n",
830 				pci_irq_vector(pdev, 1), error);
831 			goto err_free_irq;
832 		}
833 		if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM) {
834 			error = request_irq(pci_irq_vector(pdev, 2),
835 					    vmci_interrupt_dma_datagram,
836 					    0, KBUILD_MODNAME, vmci_dev);
837 			if (error) {
838 				dev_err(&pdev->dev,
839 					"Failed to allocate irq %u: %d\n",
840 					pci_irq_vector(pdev, 2), error);
841 				goto err_free_bm_irq;
842 			}
843 		}
844 	}
845 
846 	dev_dbg(&pdev->dev, "Registered device\n");
847 
848 	atomic_inc(&vmci_num_guest_devices);
849 
850 	/* Enable specific interrupt bits. */
851 	cmd = VMCI_IMR_DATAGRAM;
852 	if (caps_in_use & VMCI_CAPS_NOTIFICATIONS)
853 		cmd |= VMCI_IMR_NOTIFICATION;
854 	if (caps_in_use & VMCI_CAPS_DMA_DATAGRAM)
855 		cmd |= VMCI_IMR_DMA_DATAGRAM;
856 	vmci_write_reg(vmci_dev, cmd, VMCI_IMR_ADDR);
857 
858 	/* Enable interrupts. */
859 	vmci_write_reg(vmci_dev, VMCI_CONTROL_INT_ENABLE, VMCI_CONTROL_ADDR);
860 
861 	pci_set_drvdata(pdev, vmci_dev);
862 
863 	vmci_call_vsock_callback(false);
864 	return 0;
865 
866 err_free_bm_irq:
867 	free_irq(pci_irq_vector(pdev, 1), vmci_dev);
868 err_free_irq:
869 	free_irq(pci_irq_vector(pdev, 0), vmci_dev);
870 	tasklet_kill(&vmci_dev->datagram_tasklet);
871 	tasklet_kill(&vmci_dev->bm_tasklet);
872 
873 err_disable_msi:
874 	pci_free_irq_vectors(pdev);
875 
876 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
877 	if (vmci_err < VMCI_SUCCESS)
878 		dev_warn(&pdev->dev,
879 			 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
880 			 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
881 
882 err_remove_bitmap:
883 	if (vmci_dev->notification_bitmap) {
884 		vmci_write_reg(vmci_dev, VMCI_CONTROL_RESET, VMCI_CONTROL_ADDR);
885 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
886 				  vmci_dev->notification_bitmap,
887 				  vmci_dev->notification_base);
888 	}
889 
890 err_remove_vmci_dev_g:
891 	spin_lock_irq(&vmci_dev_spinlock);
892 	vmci_pdev = NULL;
893 	vmci_dev_g = NULL;
894 	spin_unlock_irq(&vmci_dev_spinlock);
895 
896 err_free_data_buffers:
897 	vmci_free_dg_buffers(vmci_dev);
898 
899 	/* The rest are managed resources and will be freed by PCI core */
900 	return error;
901 }
902 
903 static void vmci_guest_remove_device(struct pci_dev *pdev)
904 {
905 	struct vmci_guest_device *vmci_dev = pci_get_drvdata(pdev);
906 	int vmci_err;
907 
908 	dev_dbg(&pdev->dev, "Removing device\n");
909 
910 	atomic_dec(&vmci_num_guest_devices);
911 
912 	vmci_qp_guest_endpoints_exit();
913 
914 	vmci_err = vmci_event_unsubscribe(ctx_update_sub_id);
915 	if (vmci_err < VMCI_SUCCESS)
916 		dev_warn(&pdev->dev,
917 			 "Failed to unsubscribe from event (type=%d) with subscriber (ID=0x%x): %d\n",
918 			 VMCI_EVENT_CTX_ID_UPDATE, ctx_update_sub_id, vmci_err);
919 
920 	spin_lock_irq(&vmci_dev_spinlock);
921 	vmci_dev_g = NULL;
922 	vmci_pdev = NULL;
923 	spin_unlock_irq(&vmci_dev_spinlock);
924 
925 	dev_dbg(&pdev->dev, "Resetting vmci device\n");
926 	vmci_write_reg(vmci_dev, VMCI_CONTROL_RESET, VMCI_CONTROL_ADDR);
927 
928 	/*
929 	 * Free IRQ and then disable MSI/MSI-X as appropriate.  For
930 	 * MSI-X, we might have multiple vectors, each with their own
931 	 * IRQ, which we must free too.
932 	 */
933 	if (vmci_dev->exclusive_vectors) {
934 		free_irq(pci_irq_vector(pdev, 1), vmci_dev);
935 		if (vmci_dev->mmio_base != NULL)
936 			free_irq(pci_irq_vector(pdev, 2), vmci_dev);
937 	}
938 	free_irq(pci_irq_vector(pdev, 0), vmci_dev);
939 	pci_free_irq_vectors(pdev);
940 
941 	tasklet_kill(&vmci_dev->datagram_tasklet);
942 	tasklet_kill(&vmci_dev->bm_tasklet);
943 
944 	if (vmci_dev->notification_bitmap) {
945 		/*
946 		 * The device reset above cleared the bitmap state of the
947 		 * device, so we can safely free it here.
948 		 */
949 
950 		dma_free_coherent(&pdev->dev, PAGE_SIZE,
951 				  vmci_dev->notification_bitmap,
952 				  vmci_dev->notification_base);
953 	}
954 
955 	vmci_free_dg_buffers(vmci_dev);
956 
957 	if (vmci_dev->mmio_base != NULL)
958 		pci_iounmap(pdev, vmci_dev->mmio_base);
959 
960 	/* The rest are managed resources and will be freed by PCI core */
961 }
962 
963 static const struct pci_device_id vmci_ids[] = {
964 	{ PCI_DEVICE(PCI_VENDOR_ID_VMWARE, PCI_DEVICE_ID_VMWARE_VMCI), },
965 	{ 0 },
966 };
967 MODULE_DEVICE_TABLE(pci, vmci_ids);
968 
969 static struct pci_driver vmci_guest_driver = {
970 	.name		= KBUILD_MODNAME,
971 	.id_table	= vmci_ids,
972 	.probe		= vmci_guest_probe_device,
973 	.remove		= vmci_guest_remove_device,
974 };
975 
976 int __init vmci_guest_init(void)
977 {
978 	return pci_register_driver(&vmci_guest_driver);
979 }
980 
981 void __exit vmci_guest_exit(void)
982 {
983 	pci_unregister_driver(&vmci_guest_driver);
984 }
985